[BUG] cuvsDatasetMakePadded rejects an already-aligned device tensor

[imotov]

Describe the bug

cuvsDatasetMakePadded fails when handed a device tensor whose rows already sit at CAGRA's required stride. The refusal makes sense since it avoids a redundant copy. However, it is not documented in the API, and the C API exposes no way to predict it.

Steps/Code to reproduce bug

1. Allocate a contiguous float32 device matrix whose dimension is a multiple of 4 — e.g. 64 rows × 128 dims. (Contiguous means rowStride == columns, which for 4-byte elements is already the 16-byte-aligned width CAGRA requires.)
2. Wrap it in a DLManagedTensor with device type kDLCUDA.
3. Call cuvsDatasetMakePadded(res, tensor, CUVS_DATASET_MEM_TYPE_DEVICE, &out).
4. Repeat with 127 dims instead of 128 — it succeeds. The boundary is dim % 4 == 0 for float32.

Expected behavior

The dataset contents are copied into newly allocated padded storage, as documented at

cuvs/c/include/cuvs/core/dataset.h

Lines 62 to 63 in 2140532

	* The source tensor may reside in host- or device-accessible memory. Its contents are copied into
	* newly allocated padded storage in `target_mem_type`.

Actual behavior

CUVS_ERROR, with cuvsGetLastErrorText() returning:

RAFT failure at cpp/include/cuvs/neighbors/common.hpp:1162:
source is device and stride is already correct.
Use make_device_padded_dataset_view() to get a view instead.

Environment details

Not environment specific. Present on main and release/26.08:

cuvs/cpp/include/cuvs/neighbors/common.hpp

Lines 1161 to 1166 in 2140532

	if (device_src && src_stride == target_stride) {
	RAFT_EXPECTS(false,
	"source is device and stride is already correct. "
	"Use %s() to get a view instead.",
	view_factory_name);
	}

Additional context

The Python workaround is probably the clearest signal that the API is missing something:

cuvs/python/cuvs/cuvs/common/dataset.pyx

Lines 73 to 89 in 2140532

	cdef cuvsError_t status = cuvsDatasetMakePadded(
	res,
	dataset_dlpack,
	CUVS_DATASET_MEM_TYPE_DEVICE,
	&padded.dataset
	)
	if status == cuvsError_t.CUVS_SUCCESS:
	return padded
	err = get_last_error_text() or ""
	if "stride is already correct" not in err:
	check_cuvs(status)
	check_cuvs(cuvsDatasetMakePaddedView(
	res,
	dataset_dlpack,
	&padded.dataset
	))
	return padded

It matches on the error text, so any rewording of that message turns the workaround into a crash. Rust, Go, and Java have no equivalent workaround and propagate the failure.

Might be related: #2402 (Dataset API C++, C, and language wrapper follow-up), #2394 (Remove dataset views from language wrappers).